6.6. Example: revisiting handwriting classification

Consider the following hypothetical situation. Your manager comes to you and says, “That handwriting recognition program you made is great, but it takes up too much memory and customers can’t download our application over the air. (At the time of writing there’s a 10 MB limit on certain applications downloaded over the air. I’m sure this will be laughable at some point in the future.) We need you to keep the same performance with less memory used. I told the CEO you’d have this ready in a week. How long will it take?” I’m not sure how you’d respond, but if you wanted to comply with their request, you could consider using support vector machines. The k-Nearest Neighbors algorithm used in chapter 2 works well, but you have to carry around all the training examples. With support vector machines, you can carry around far fewer examples (only your support vectors) and achieve comparable performance.

Example: digit recognition with SVMs Collect: Text file provided. Prepare: Create vectors from the binary images. Analyze: Visually inspect the image vectors. Train: Run the SMO algorithm with two different kernels and different settings for the radial bias kernel. Test: Write a function to test the different kernels and calculate the error rate. Use: A full application of image recognition requires some image processing, which we won’t get into.